The public release of DeepSeek R1 in January 2025 changed the AI landscape enormously. But, today, I’m going to talk about the next step in DeepSeek’s evolution, and perhaps the next big leap forward in AI technology as we know it: DeepSeek R2.
Recent Advancements of DeepSeek
This has definitely been going under the radar somewhat, but DeepSeek’s research team recently hosted an even called Open-Source Week. During that event, the DeepSeek team released five open-source repos to show the recent progress and AI advancements they’ve been making.
The three most notable of those tools are:
- FlashMLA: This is an efficient MLA decoding kernel for Hopper GPUs.
- DeepEP: The first ever communication library for MoE model training and inference.
- DeepGEMM: An FP8 GEMM library that works with both dense and MoE GEMMs.
So, what does this all actually mean?
Well, all of these tools are designed to make it easier and more efficient to develop and streamline future AI models and products. And that leads me neatly onto the main topic of discussion: DeepSeek R2.
DeepSeek R2: The Next Evolution in AI
DeepSeek’s Open-Source Week event and the products showcased there all tie neatly into the upcoming release of the company’s much-hyped Reasoning 2 or R2 model, which is set to be a major upgrade over the previous DeepSeek R1 model.
For anyone out there who hasn’t been following DeepSeek-related news and releases, I’ll break down the basics:
DeepSeek R1 was developed in China launched in early 2025. It was a major event in the AI sphere, and it partially led to the selling of more than $1 trillion from the global equities market. Why? Because R1 was such an advanced, revolutionary model, more than capable of competing with the best AI models from the big Western brands.
Testing showed it was, in many areas, equivalent to or better than the likes of:
- Claude
- GPT-4o
- OpenAI-o1-mini
- And many others
It managed to do all this while being open-source and free to access. It was groundbreaking at the time, and I was one of many AI enthusiasts across the globe to be blown away by the story: a small Chinese startup had managed to achieve the same level of technological excellence as huge multibillion-dollar brands like OpenAI.
But it didn’t take long after R1’s release for me and many others to wonder, “What’s next?”
Well, a Reuters report recently came out, revealing that DeepSeek’s development team is in a hurry to get the R2 model rolled out by May. So we might be getting it very soon, and I can tell you for certain: this new model is going to change the AI landscape even more dramatically than R1 did.
The Impact of R2
Here’s what I expect from R2:
- It will be extremely cheap, just like R1
- It will match or even outperform leading models like OpenAI’s o3-full or o3-high
- It will outperform almost all other models on the market
And it’s coming really soon, if the latest reports are to be believed. The Hangzhou-based team working on DeepSeek seems really eager to get it out, and my personal prediction is that we might see it release just after the Chinese Qingming Festival holiday, which is on the 6th of May, so keep an eye on your calendars.
The DeepSeek team is also said to be focusing on producing superior code than before and better performance in a range of languages, not just English and Chinese, which were the main focuses and default options with R1. This makes sense, as DeepSeek R1 proved popular all over the world, and people will want to interact with R2 in Spanish, French, Russian, etc.
And we only have to look back at the impact R1 had to predict how R2 might affect the global markets. R1’s release led to a massive dip in the stock market, with the likes of NVIDIA stock, crypto, and global equities all dropping in the days following its release. That wowed me at the time, because we’d almost never seen anything like it before.
Why did that happen, exactly? Well, partly because R1 was built using less powerful, older NVIDIA chips, not the latest ones, so it made NVIDIA’s latest tech less valuable and impressive from that standpoint. It also brought big competition to the major Western AI tech giants, leading to lower valuations for them and their products.
The launch of R2 could have similar impact. It could disrupt the dominance of the big AI firms, like OpenAI and Google, shaking up the landscape yet again and proving that you don’t necessarily need to use the absolute cutting-edge, latest and greatest chips and other pieces of hardware to make incredible AI models.
DeepSeek’s Secret to Success
By this point, you might be wondering “What is the secret behind DeepSeek’s incredible success story?” Well, I think I’ve got the answer. The key to DeepSeek’s success was in how they rapidly innovated in AI development, daring to think outside the box and do things differently from the big, established firms.
Innovation in Technology
A big part of this was a heavy investment in computing. DeepSeek’s parent company poured a lot of money into purchasing computing hardware, like state-of-the-art supercomputing clusters, like Firefly, which uses thousands of NVIDIA 800 chips at much lower cost than the latest chips, allowing DeepSeek to build amazing AI on a relatively small budget.
In fact, there’s quite an interesting story behind that. The company behind DeepSeek spent around 1.2 billion yuan on two of these supercomputing AI clusters in 2020 and 2021. Those clusters were made up of around 10,000 NVIDIA 800 chips, and these purchases attracted the attention of some big regulatory bodies in China.
Those bodies asked DeepSeek why they were buying so many chips, and the team had to explain what they were working on. In the end, they got the “all clear” from the regulatory authorities, which allowed them to push on. Meanwhile, Western AI firms are using even more – up to 50,000 – high-end NVIDIA chips that have literally been banned for export to China.
So, DeepSeek had to cope with lesser quality chips, but still made it work, focusing on efficiency over peak performance, and extracting as much value as possible out of the components available. That also means lower computational costs and overheads, which are passed onto the end users – people like me and you.
Setting the Standard
Once DeepSeek had that initial infrastructure established, they were able to push forward with their innovations, scaling up their work and cost-efficiently develop different AI models and systems. They even reached a point where their model is just 4.5 points away on benchmark tests from being the official best in the world.
That is jaw-dropping to me, and it gets me even more excited about R2. Because if R1 is so close to being the best in the business, R2 will almost certainly push far higher and achieve so much more. It’s going to be very exciting to see when R2 releases how impressive it is, how it performs against other models, and how other firms will have to play catch up with DeepSeek.
Fair Pricing
There’s also the pricing factor. DeepSeek was available and accessible at dramatically cheaper price points than any other major AI model on the market. I compared pricing charts at the time of R1’s release, and the difference was staggering. And many people quickly abandoned their OpenAI subscriptions and flocked to DeepSeek when they saw how good and cheap it was.
We even saw OpenAI having to quickly reduce its pricing plans, because DeepSeek was up to 40 times cheaper than what OpenAI was charging at the time, which still blows me away when I think back to it.
Of course, we don’t know much about R2’s pricing just yet, but I don’t expect DeepSeek to change tactics too much with the new model. It should still be an affordable option, giving as many people as possible access to high-end AI technology without forcing them to spend huge amounts on monthly fees.
Plus, with DeepSeek releasing these new tools at its Open-Source event, the bar-of-entry to AI development just got even lower and more affordable. So it’s possible that R2 might be even cheaper than the original R1 model, which would truly send seismic waves throughout the entire AI world and beyond.
Get Ready for the DeepSeek R2 Release
Overall, I’m really excited for the upcoming R2 release from DeepSeek. All the signs are positive, suggesting that this new open-source model will be even better than not just R1, but any other AI model we’ve seen so far. It’s going to be faster, smarter, and superior in every way, and we might even get a better price for it, too. Stay tuned for further updates.